Another to_cfg crash

[oflatt]

@main() {
    i: int = const 1;
    jmp .loop;
.loop:
    max: int = const 10;
    cond: bool = lt i max;
    br cond .t .f;
.t:
    res: int = const 7;
    jmp .loopbottom;
.f:
    res: int = const 8;
    jmp .loopbottom;
.loopbottom:
    one: int = const 1;
    i: int = add i one;
    br cond .loop .exit;
.exit:
    print res;
}

[oflatt]

Looks like the original RVSDG generated from the bril program is referring to a non-existant argument

[ezrosent]

This one is interesting: Basically, what's happening is that res is not defined at the start of the loop, but because it's always assigned inside the loop and the loop itself runs at least once, it's safe to use it as an output.

Lots of ways to handle this: "undef" variables? Adding special output-only ports to thetas?

[oflatt]

That is interesting!
Alternative proposal: Can we disallow these programs? I'd like to throw an error for undef variables